[PowerPC] Add fast-isel branch and compare selection.

First chunk of actual fast-isel selection code.  This handles direct
and indirect branches, as well as feeding compares for direct
branches.  PPCFastISel::PPCEmitIntExt() is just roughed in and will be
expanded in a future patch.  This also corrects a problem with
selection for constant pool entries in JIT mode or with small code
model.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189202 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/PowerPC/PPCFastISel.cpp

@@ -89,9 +89,19 @@ class PPCFastISel : public FastISel {
     virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
                                      const LoadInst *LI);
     virtual bool FastLowerArguments();
+    virtual unsigned FastEmit_i(MVT Ty, MVT RetTy, unsigned Opc, uint64_t Imm);
+
+  // Instruction selection routines.
+  private:
+    bool SelectBranch(const Instruction *I);
+    bool SelectIndirectBr(const Instruction *I);
 
   // Utility routines.
   private:
+    bool PPCEmitCmp(const Value *Src1Value, const Value *Src2Value,
+                    bool isZExt, unsigned DestReg);
+    bool PPCEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
+                           unsigned DestReg, bool IsZExt);
     unsigned PPCMaterializeFP(const ConstantFP *CFP, MVT VT);
     unsigned PPCMaterializeInt(const Constant *C, MVT VT);
     unsigned PPCMaterialize32BitInt(int64_t Imm,
@@ -106,10 +116,241 @@ class PPCFastISel : public FastISel {
 
 } // end anonymous namespace
 
+static Optional<PPC::Predicate> getComparePred(CmpInst::Predicate Pred) {
+  switch (Pred) {
+    // These are not representable with any single compare.
+    case CmpInst::FCMP_FALSE:
+    case CmpInst::FCMP_UEQ:
+    case CmpInst::FCMP_UGT:
+    case CmpInst::FCMP_UGE:
+    case CmpInst::FCMP_ULT:
+    case CmpInst::FCMP_ULE:
+    case CmpInst::FCMP_UNE:
+    case CmpInst::FCMP_TRUE:
+    default:
+      return Optional<PPC::Predicate>();
+
+    case CmpInst::FCMP_OEQ:
+    case CmpInst::ICMP_EQ:
+      return PPC::PRED_EQ;
+
+    case CmpInst::FCMP_OGT:
+    case CmpInst::ICMP_UGT:
+    case CmpInst::ICMP_SGT:
+      return PPC::PRED_GT;
+
+    case CmpInst::FCMP_OGE:
+    case CmpInst::ICMP_UGE:
+    case CmpInst::ICMP_SGE:
+      return PPC::PRED_GE;
+
+    case CmpInst::FCMP_OLT:
+    case CmpInst::ICMP_ULT:
+    case CmpInst::ICMP_SLT:
+      return PPC::PRED_LT;
+
+    case CmpInst::FCMP_OLE:
+    case CmpInst::ICMP_ULE:
+    case CmpInst::ICMP_SLE:
+      return PPC::PRED_LE;
+
+    case CmpInst::FCMP_ONE:
+    case CmpInst::ICMP_NE:
+      return PPC::PRED_NE;
+
+    case CmpInst::FCMP_ORD:
+      return PPC::PRED_NU;
+
+    case CmpInst::FCMP_UNO:
+      return PPC::PRED_UN;
+  }
+}
+
+// Attempt to fast-select a branch instruction.
+bool PPCFastISel::SelectBranch(const Instruction *I) {
+  const BranchInst *BI = cast<BranchInst>(I);
+  MachineBasicBlock *BrBB = FuncInfo.MBB;
+  MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
+  MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
+
+  // For now, just try the simplest case where it's fed by a compare.
+  if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
+    Optional<PPC::Predicate> OptPPCPred = getComparePred(CI->getPredicate());
+    if (!OptPPCPred)
+      return false;
+
+    PPC::Predicate PPCPred = OptPPCPred.getValue();
+
+    // Take advantage of fall-through opportunities.
+    if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
+      std::swap(TBB, FBB);
+      PPCPred = PPC::InvertPredicate(PPCPred);
+    }
+
+    unsigned CondReg = createResultReg(&PPC::CRRCRegClass);
+
+    if (!PPCEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned(),
+                    CondReg))
+      return false;
+
+    BuildMI(*BrBB, FuncInfo.InsertPt, DL, TII.get(PPC::BCC))
+      .addImm(PPCPred).addReg(CondReg).addMBB(TBB);
+    FastEmitBranch(FBB, DL);
+    FuncInfo.MBB->addSuccessor(TBB);
+    return true;
+
+  } else if (const ConstantInt *CI =
+             dyn_cast<ConstantInt>(BI->getCondition())) {
+    uint64_t Imm = CI->getZExtValue();
+    MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
+    FastEmitBranch(Target, DL);
+    return true;
+  }
+
+  // FIXME: ARM looks for a case where the block containing the compare
+  // has been split from the block containing the branch.  If this happens,
+  // there is a vreg available containing the result of the compare.  I'm
+  // not sure we can do much, as we've lost the predicate information with
+  // the compare instruction -- we have a 4-bit CR but don't know which bit
+  // to test here.
+  return false;
+}
+
+// Attempt to emit a compare of the two source values.  Signed and unsigned
+// comparisons are supported.  Return false if we can't handle it.
+bool PPCFastISel::PPCEmitCmp(const Value *SrcValue1, const Value *SrcValue2,
+                             bool IsZExt, unsigned DestReg) {
+  Type *Ty = SrcValue1->getType();
+  EVT SrcEVT = TLI.getValueType(Ty, true);
+  if (!SrcEVT.isSimple())
+    return false;
+  MVT SrcVT = SrcEVT.getSimpleVT();
+
+  // See if operand 2 is an immediate encodeable in the compare.
+  // FIXME: Operands are not in canonical order at -O0, so an immediate
+  // operand in position 1 is a lost opportunity for now.  We are
+  // similar to ARM in this regard.
+  long Imm = 0;
+  bool UseImm = false;
+
+  // Only 16-bit integer constants can be represented in compares for 
+  // PowerPC.  Others will be materialized into a register.
+  if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(SrcValue2)) {
+    if (SrcVT == MVT::i64 || SrcVT == MVT::i32 || SrcVT == MVT::i16 ||
+        SrcVT == MVT::i8 || SrcVT == MVT::i1) {
+      const APInt &CIVal = ConstInt->getValue();
+      Imm = (IsZExt) ? (long)CIVal.getZExtValue() : (long)CIVal.getSExtValue();
+      if ((IsZExt && isUInt<16>(Imm)) || (!IsZExt && isInt<16>(Imm)))
+        UseImm = true;
+    }
+  }
+
+  unsigned CmpOpc;
+  bool NeedsExt = false;
+  switch (SrcVT.SimpleTy) {
+    default: return false;
+    case MVT::f32:
+      CmpOpc = PPC::FCMPUS;
+      break;
+    case MVT::f64:
+      CmpOpc = PPC::FCMPUD;
+      break;
+    case MVT::i1:
+    case MVT::i8:
+    case MVT::i16:
+      NeedsExt = true;
+      // Intentional fall-through.
+    case MVT::i32:
+      if (!UseImm)
+        CmpOpc = IsZExt ? PPC::CMPLW : PPC::CMPW;
+      else
+        CmpOpc = IsZExt ? PPC::CMPLWI : PPC::CMPWI;
+      break;
+    case MVT::i64:
+      if (!UseImm)
+        CmpOpc = IsZExt ? PPC::CMPLD : PPC::CMPD;
+      else
+        CmpOpc = IsZExt ? PPC::CMPLDI : PPC::CMPDI;
+      break;
+  }
+
+  unsigned SrcReg1 = getRegForValue(SrcValue1);
+  if (SrcReg1 == 0)
+    return false;
+
+  unsigned SrcReg2 = 0;
+  if (!UseImm) {
+    SrcReg2 = getRegForValue(SrcValue2);
+    if (SrcReg2 == 0)
+      return false;
+  }
+
+  if (NeedsExt) {
+    unsigned ExtReg = createResultReg(&PPC::GPRCRegClass);
+    if (!PPCEmitIntExt(SrcVT, SrcReg1, MVT::i32, ExtReg, IsZExt))
+      return false;
+    SrcReg1 = ExtReg;
+
+    if (!UseImm) {
+      unsigned ExtReg = createResultReg(&PPC::GPRCRegClass);
+      if (!PPCEmitIntExt(SrcVT, SrcReg2, MVT::i32, ExtReg, IsZExt))
+        return false;
+      SrcReg2 = ExtReg;
+    }
+  }
+
+  if (!UseImm)
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc), DestReg)
+      .addReg(SrcReg1).addReg(SrcReg2);
+  else
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc), DestReg)
+      .addReg(SrcReg1).addImm(Imm);
+
+  return true;
+}
+
+// Attempt to emit an integer extend of SrcReg into DestReg.  Both
+// signed and zero extensions are supported.  Return false if we
+// can't handle it.  Not yet implemented.
+bool PPCFastISel::PPCEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
+                                unsigned DestReg, bool IsZExt) {
+  return (SrcVT == MVT::i8 && SrcReg && DestVT == MVT::i8 && DestReg
+          && IsZExt && false);
+}
+
+// Attempt to fast-select an indirect branch instruction.
+bool PPCFastISel::SelectIndirectBr(const Instruction *I) {
+  unsigned AddrReg = getRegForValue(I->getOperand(0));
+  if (AddrReg == 0)
+    return false;
+
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::MTCTR8))
+    .addReg(AddrReg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::BCTR8));
+
+  const IndirectBrInst *IB = cast<IndirectBrInst>(I);
+  for (unsigned i = 0, e = IB->getNumSuccessors(); i != e; ++i)
+    FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[IB->getSuccessor(i)]);
+
+  return true;
+}
+
 // Attempt to fast-select an instruction that wasn't handled by
-// the table-generated machinery.  TBD.
+// the table-generated machinery.
 bool PPCFastISel::TargetSelectInstruction(const Instruction *I) {
-  return I && false;
+
+  switch (I->getOpcode()) {
+    case Instruction::Br:
+      return SelectBranch(I);
+    case Instruction::IndirectBr:
+      return SelectIndirectBr(I);
+    // Here add other flavors of Instruction::XXX that automated
+    // cases don't catch.  For example, switches are terminators
+    // that aren't yet handled.
+    default:
+      break;
+  }
+  return false;
 }
 
 // Materialize a floating-point constant into a register, and return
@@ -131,15 +372,18 @@ unsigned PPCFastISel::PPCMaterializeFP(const ConstantFP *CFP, MVT VT) {
       MachinePointerInfo::getConstantPool(), MachineMemOperand::MOLoad,
       (VT == MVT::f32) ? 4 : 8, Align);
 
-  // For small code model, generate a LDtocCPT.
-  if (CModel == CodeModel::Small || CModel == CodeModel::JITDefault)
+  unsigned Opc = (VT == MVT::f32) ? PPC::LFS : PPC::LFD;
+  unsigned TmpReg = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
+
+  // For small code model, generate a LF[SD](0, LDtocCPT(Idx, X2)).
+  if (CModel == CodeModel::Small || CModel == CodeModel::JITDefault) {
     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::LDtocCPT),
-            DestReg)
-      .addConstantPoolIndex(Idx).addReg(PPC::X2).addMemOperand(MMO);
-  else {
+            TmpReg)
+      .addConstantPoolIndex(Idx).addReg(PPC::X2);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
+      .addImm(0).addReg(TmpReg).addMemOperand(MMO);
+  } else {
     // Otherwise we generate LF[SD](Idx[lo], ADDIStocHA(X2, Idx)).
-    unsigned Opc = (VT == MVT::f32) ? PPC::LFS : PPC::LFD;
-    unsigned TmpReg = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::ADDIStocHA),
             TmpReg).addReg(PPC::X2).addConstantPoolIndex(Idx);
     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
@@ -312,6 +556,25 @@ bool PPCFastISel::FastLowerArguments() {
   return false;
 }
 
+// Handle materializing integer constants into a register.  This is not
+// automatically generated for PowerPC, so must be explicitly created here.
+unsigned PPCFastISel::FastEmit_i(MVT Ty, MVT VT, unsigned Opc, uint64_t Imm) {
+  
+  if (Opc != ISD::Constant)
+    return 0;
+
+  if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16 &&
+      VT != MVT::i8 && VT != MVT::i1) 
+    return 0;
+
+  const TargetRegisterClass *RC = ((VT == MVT::i64) ? &PPC::G8RCRegClass :
+                                   &PPC::GPRCRegClass);
+  if (VT == MVT::i64)
+    return PPCMaterialize64BitInt(Imm, RC);
+  else
+    return PPCMaterialize32BitInt(Imm, RC);
+}
+
 namespace llvm {
   // Create the fast instruction selector for PowerPC64 ELF.
   FastISel *PPC::createFastISel(FunctionLoweringInfo &FuncInfo,

test/CodeGen/PowerPC/fast-isel-br-const.ll

@@ -0,0 +1,43 @@
+; RUN: llc < %s -O0 -verify-machineinstrs -fast-isel-abort -mtriple=powerpc64-unknown-linux-gnu -mcpu=pwr7 | FileCheck %s --check-prefix=ELF64
+
+define i32 @t1(i32 %a, i32 %b) nounwind uwtable ssp {
+entry:
+; ELF64: t1
+  %x = add i32 %a, %b  
+  br i1 1, label %if.then, label %if.else
+; ELF64-NOT: b {{\.?}}LBB0_1
+
+if.then:                                          ; preds = %entry
+  call void @foo1()
+  br label %if.end7
+
+if.else:                                          ; preds = %entry
+  br i1 0, label %if.then2, label %if.else3
+; ELF64: b {{\.?}}LBB0_4
+
+if.then2:                                         ; preds = %if.else
+  call void @foo2()
+  br label %if.end6
+
+if.else3:                                         ; preds = %if.else
+  %y = sub i32 %a, %b
+  br i1 1, label %if.then5, label %if.end
+; ELF64-NOT: b {{\.?}}LBB0_5
+
+if.then5:                                         ; preds = %if.else3
+  call void @foo1()
+  br label %if.end
+
+if.end:                                           ; preds = %if.then5, %if.else3
+  br label %if.end6
+
+if.end6:                                          ; preds = %if.end, %if.then2
+  br label %if.end7
+
+if.end7:                                          ; preds = %if.end6, %if.then
+  ret i32 0
+}
+
+declare void @foo1()
+
+declare void @foo2()

test/CodeGen/PowerPC/fast-isel-indirectbr.ll

@@ -0,0 +1,15 @@
+; RUN: llc < %s -O0 -verify-machineinstrs -fast-isel-abort -mtriple=powerpc64-unknown-linux-gnu -mcpu=pwr7 | FileCheck %s --check-prefix=ELF64
+
+define void @t1(i8* %x) {
+entry:
+; ELF64: t1
+  br label %L0
+
+L0:
+  br label %L1
+
+L1:
+  indirectbr i8* %x, [ label %L0, label %L1 ]
+; ELF64: mtctr 3
+; ELF64: bctr
+}